TSR-GAN: Generative Adversarial Networks for Traffic State Reconstruction with Time Space Diagrams

Abstract

A Time Space Diagram (TSD) plays an important role in transportation research and practice due to its capability to exhibit traffic dynamics in time and space. Based on TSDs, this paper aims to reconstruct the traffic spatio-temporal state with the aid of Generative Adversarial Networks (GANs). By mining traffic state correlations and traffic pattern similarities between lanes with or without sufficient observations, the proposed Traffic State Reconstruction GAN (TSR-GAN) model can well estimate the traffic states for road segments with a strong learning capability. Specifically, the traffic states of lanes are converted to TSDs, in which the color represents the values of traffic variables (e.g., speed or density). The TSDs of lanes with or without sufficient data are utilized to train the proposed TSR-GAN model. The fine-tuned TSR-GAN model reconstructs traffic states for road segments with deficient sensor coverage by restoring the high-resolution TSD from its low-resolution observation. With trajectory datasets from Next Generation Simulation (NGSIM), this paper verifies the performance of the TSR-GAN model by estimating travel time via the reconstructed TSDs. Numerical results demonstrate that the proposed model possesses a desirable generalization and transferability, demonstrating the promise of reconstructing traffic states under various conditions.